AI Filmmaking Part 1 | Video gen with LTX 2.3: You're doing it wrong.
See also ComfyUI Tutorial Series_ Ep01 - Introduction and Installation
and Claude Code + OpenRouter = Free UNLIMITED Coding (No RAM Needed)
1,761 views Jun 4, 2026
This is part 1 of a series I am making covering the finer details of AI Filmmaking using 100% local, open source, free models & tools. Link to workflow (LTX 2.3 Seed Hunting Multiroll): https://civitai.com/models/2676452
Transcript
0:00 · Hi, I'm Tristan and I use 100% local open-source models and tools to make short [music] films.
0:06 · If you haven't yet, please check out my previous work here on the channel.
0:10 · Going forward, I will be focused on making extremely detailed guides about my approach to local model AI filmmaking. From creating consistent characters that can be placed and posed for shot compositions to final production editing strategies that fix many common issues we see with AI video today. Most importantly, everything I will cover and use is completely free.
0:31 · So, if this interests you, please like and subscribe and leave a comment if I've helped you in any way.
0:37 · Really helps as a new creator in this already extremely niche space.
0:41 · We all know LTX is abysmal at prompt adherence. People look to a plethora of prompting guides to solve this glaring issue, but only end up becoming more frustrated as LTX continuously [music] refuses to obey even simple requests sometimes. Prompting skill only helps so much and well, I think Dezra says it best.
1:01 · You're using LTX wrong.
1:02 · [music] Part one of this tutorial series covers the most important reason why you're not seed hunting. My discovery from hundreds of hours of experience making short films using LTX is that seed hunting is the one true path to success with this model, not banging your head against the wall trying to prompt better.
1:26 · To seed hunt, you need to be able to quickly see the outcome of many different seeds and then be able to pick one to carry over and gen in full resolution.
1:36 · So, then the question remains, how do we do this? How do we window shop for a good seed?
1:41 · What you're seeing right now is my custom workflow that generates to 1080p in three separate stages. The first stage generates four low-resolution [music] samples, each with a unique seed that as you can see produces many different outcomes.
1:56 · This stage takes very little time, often just 20 to 60 seconds to generate all four sample gens, depending on your hardware, resolution, and frame length, of course. Next, I toggle finish mode, which enables the second and third stages [music] and takes the selected sample latent, in this case number three, onwards towards finalization.
2:17 · The second stage will begin [music] to fine-tune the visuals in motion, while always looking at the original input image as reference. When each stage finishes, you can always cancel if you don't see a result you like, and simply click the reroll button on the specific stage to redo it without having to gen everything over from the beginning. In this case, I like what I see, so I let it continue to stage three.
2:41 · This process is similar to gold panning in a river. You take a bunch of mud and rocks, shake them around in your sieve until all the unwanted crap is gone, and all that remains is gold. Or, in this metaphor's case, a finalized 1080p gen that is both visually coherent and has your desired motion.
2:57 · So now, using this seed hunting workflow, what are the top tricks and strategies to get great results out of LTX?
3:05 · Number one, supply your own custom audio, especially for dialogue. Nearly every shot in any of my short films will have used custom audio, either supplied to LTX during the gen process or added in post during the video editing phase. This is because you have very little control over the outcome of LTX's generative audio. It will add music randomly, [music] it's low quality, and you can't achieve consistent voices between gens.
3:32 · Oh, who am I kidding? You're using LTX WRONG!
3:37 · IT'S FOR ALL intents and purposes garbage, only useful for the occasional otherwise tedious task like adding footsteps to a character walking. But while LTX's native generative audio is garbo trash, LTX's internal understanding of sound is extremely intelligent. You can supply a line of dialogue as custom audio and prompt only that the character speaking and it will lip sync the dialogue perfectly without even prompting what the character said.
4:05 · Furthermore, sound effects in supplied audio will be timed to actions in the gen itself such as this fireball hitting the tree. Reroll your seed for iterative variations at each stage allowing you to narrow in and fine-tune your outcome.
4:24 · After you've got your stage one samples and picked a promising latent to take to the next stage, so long as you don't change any values or settings that would affect those stage one latents, they will remain in memory and available to be chosen for finish mode processing.
4:42 · If you decide your selected stage one latent wasn't a good choice, you can change your selection as many times as you like proceeding to stage two without having to regen anything prior. And if something looks off about a resulting preview, simply cancel the run and click the manual random seed button for the stage and it'll generate a new version. In this case, I'm rerolling stage two to see if I can get better motion.
5:13 · The same process can be done for stage three. If you don't like your final result but like what you see from stage two's preview, reroll it with the manual random seed button. Stage three takes the longest because it's 1080p but sometimes it's worth the time to perfect the motion or visual quality of your gen, especially when it comes to fast motion or lip syncing. This ability to reroll set seeds at each stage of processing gives you a huge degree of control over how closely your final gen matches your desired goal.
5:46 · The The path to frustration is to demand too much from LTX in a single gen. Whether that be a long 15-second tracking shot or a complicated shorter shot with lots of motion. LTX can give you bad results after bad results even with the power of seed hunting on your side.
6:05 · The solution is to break up your gen into smaller actions using video extension. That way LTX is never thinking too hard and ruining your gen in one way or another. Thankfully, [music] LTX natively understands how to extend video and I've implemented the feature into this workflow.
6:27 · While it's often [music] best to use first frame last frame workflows for video extension, and that would be [music] starting with the video frames as key frame input and ending with the last frame that [music] guides the gen to a coherent destination and reminds the model what your characters and scene looks like. That specific subject will be covered in a future video.
6:48 · For now, we're just going to extend the trash can video we made in the last section.
6:53 · [music] Of course, you could have done this all in one 10-second gen, but the more you ask LTX to do in a single prompt, [music] the more that can go wrong with the motion or visual fidelity and the longer it takes you to find out you have an unusable result. It will save you time in the end to simply break complicated [music] shots up into multiple parts and merge them together in video editing. This process is theoretically similar to what we were doing with rerolls in the last section.
7:24 · Locking in motion we like and fine-tuning from there. I hope this workflow and methodology resonates with you. To recap, LTX does not care about your prompt as much as models like 1 2.2 And thus, it requires a different approach. Each seed produces radically different results, and the path to success is being able to quickly find that golden seed that matches the motion you were looking for.
7:55 · Secondly, while generative audio is a nice feature, it is underbaked and often unusable. Meaning, for serious filmmaking, being able to supply high-quality custom audio for your gens is crucial. Third, being able to reroll for detail / motion variations at each stage offers fine-tuning that narrows in on your perfect shot without having to regen from scratch every time.
8:20 · Then lastly, building on shots by extending them rather than throwing darts at the board and hoping they all hit the bull's-eye at once is the best way to stay sane with this model. So, stay tuned for more insights and tips into AI filmmaking, such as covering that first frame last frame workflow I mentioned earlier.
8:41 · Also, topics like V2V editing, video outpainting to create epic wide shots, consistent character voice generation, and video editing tricks to cover up a lot of those common AI problems. I do offer one-on-one online coaching at a pretty cheap hourly rate, and I do have a Patreon if you wish to support what I do. Info in the description below. Thanks, and take care.